A novel hybrid recurrent convolutional network for surrogate modeling of history matching and uncertainty quantification

Automatic history matching (AHM) has been widely studied in petroleum engineering due to it can provide reliable numerical models for reservoir development and management. However, AHM is still a challenging problem because it usually involves running a great deal of time-consuming numerical simulations during the solving process. To address this issue, this article studies a hybrid recurrent convolutional network (HRCN) model for surrogate modeling of numerical simulation used in AHM. The HRCN model is end-to-end trainable for predicting the well production data of high-dimensional parameter fields. In HRCN, a convolutional neural network (CNN) is first developed to learn the high-level spatial feature representations of the input parameter fields. Following that, a recurrent neural network (RNN) is constructed with the purpose of modeling complex temporal dynamics and predicting the production data. In addition, given that the fluctuations of production data are influenced by well control measures, the well control parameters are used as auxiliary inputs of RNN. Moreover, the proposed surrogate model is incorporated into a multimodal estimation of distribution algorithm (MEDA) to formulate a novel surrogate-based AHM workflow. The numerical studies performed on a 2D and a 3D reservoir model illustrate the performance of the proposed surrogate model and history matching workflow. Compared with the MEDA using only numerical simulations, the surrogate-based AHM workflow significantly reduces the computational cost.

Automated summary

This article studies a hybrid recurrent convolutional network (HRCN) model for surrogate modeling of numerical simulation used in automatic history matching (AHM). The HRCN model combines convolutional neural network (CNN) and recurrent neural network (RNN) to learn spatial and temporal features for predicting well production data and reduces computational cost.